Join us at TensorFlow World, Oct 28-31. Use code TF20 for 20% off select passes. Register now

Post-training quantization

Post-training quantization includes general techniques to reduce CPU and hardware accelerator latency, processing, power, and model size with little degradation in model accuracy. These techniques can be performed on an already-trained float TensorFlow model and applied during TensorFlow Lite conversion. These techniques are enabled as options in the TensorFlow Lite converter:

Quantizing weights

Weights can be converted to types with reduced precision, such as 16 bit floats or 8 bit integers. We generally recommend 16-bit floats for GPU acceleration and 8-bit integer for CPU execution.

For example, here is how to specify 8 bit integer weight quantization:

import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
tflite_quant_model = converter.convert()

At inference, the most critically intensive parts are computed with 8 bits instead of floating point. There is some inference-time performance overhead, relative to quantizing both weights and activations below.

For more information, see the TensorFlow Lite post-training quantization guide.

Full integer quantization of weights and activations

Improve latency, processing, and power usage, and get access to integer-only hardware accelerators by making sure both weights and activations are quantized. This requires a small representative data set.

import tensorflow as tf

def representative_dataset_gen():
  for _ in range(num_calibration_steps):
    # Get sample input data as a numpy array in a method of your choosing.
    yield [input]

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
tflite_quant_model = converter.convert()

The resulting model will still take float input and output for convenience.

For more information, see the TensorFlow Lite post-training quantization guide.